Robot Learning for Manipulation of Deformable Linear Objects

Licentiatavhandling, 2021

Deformable Object Manipulation (DOM) is a challenging problem in robotics. Until recently there has been limited research on the subject, with most robotic manipulation methods being developed for rigid objects. Part of the challenge in DOM is that non-rigid objects require solutions capable of generalizing to changes in shape and mechanical properties. Recently, Machine Learning (ML) has been proven successful in other fields where generalization is important such as computer vision, thus encouraging the application of ML to robotics as well. Notably, Reinforcement Learning (RL) has shown promise in finding control policies for manipulation of rigid objects. However, RL requires large amounts of data that are better satisfied in simulation while deformable objects are inherently more difficult to model and simulate.

This thesis presents ReForm, a simulation sandbox for robotic manipulation of Deformable Linear Objects (DLOs) such as cables, ropes, and wires. DLO manipulation is an interesting problem for a variety of applications throughout manufacturing, agriculture, and medicine. Currently, this sandbox includes six shape control tasks, which are classified as explicit when a precise shape is to be achieved, or implicit when the deformation is just a consequence of a more abstract goal, e.g. wrapping a DLO around another object. The proposed simulation environments aim to facilitate comparison and reproducibility of robot learning research. To that end, an RL algorithm is tested on each simulated task providing initial benchmarking results. ReForm is one of three concurrent frameworks to first support DOM problems.

This thesis also addresses the problem of DLO state representation for an explicit shape control problem. Moreover, the effects of elastoplastic properties on the RL reward definition are investigated. From a control perspective, DLOs with these properties are particularly challenging to manipulate due to their nonlinear behavior, acting elastic up to a yield point after which they become permanently deformed. A low-dimensional representation from discrete differential geometry is proposed, offering more descriptive shape information than a simple point-cloud while avoiding the need for curve fitting. Empirical results show that this representation leads to a better goal description in the presence of elastoplasticity, preventing the RL algorithm from converging to local minima which correspond to incorrect shapes of the DLO.